EAGER: Fairness-Aware Personalized Recommendations
NSF IIS: 1841138, August 2018 to July 2020 (expected)
Project Goals: The goal of this project is to create effective information curator recommendation models that can be personalized for individual users, while maintaining important fairness properties. The advances in uncovering information curators at scale, reliably connecting users to appropriate curators, and ensuring fairness-preserving properties of such curators are critical for trustworthy information supporting an informed populace. By bringing these research advances, datasets, and toolkits to the wider research community, this project can spur additional advances from complementary efforts by other researchers.
Research Challenges: A key challenge for personalized curator recommendation is tackling sparsity while carefully modeling curators in complex, noisy, and heterogeneous environments. Compounding this challenge, most current access to information curators is mediated by centralized platforms (like search engines, social networks, and traditional news media), meaning that personal preferences may not align with the goals of these platforms, leading to potentially biased (or even limited) access to curators. A key question is how to maintain fairness properties in curator recommendation.
Broader Impacts: The successful outcome of this project will lead to research advances that can positively impact existing web and social media platforms, as well as provide a theoretical foundation for future advances in information curation recommendation. The advances in uncovering information curators at scale, reliably connecting users to appropriate curators, and ensuring fairness-preserving properties of such curators are critical for a trustworthy information diet supporting an informed populace. This project will develop new classroom materials, new outreach efforts, and new broadening participation workshops and seminars. Together, these efforts will integrate the new knowledge developed as part of the research plan through investments in undergraduate and graduate students, and through course enhancements and research training.
Current Results: In the first year of the project, our research team made a number of strides in improving recommendation, guided by our proposed research effort. Concretely, we have focused on: 1) new neural models of recommendation; and 2) new fairness-aware approaches for recommendation.
- In our first research thrust, our goal is to create new neural models for personalized recommendation. Neural models promise potentially more flexibility in model design, added nonlinearity through activations, and improved performance relative to traditional approaches. Indeed, many recent efforts have highlighted this potential. For this thrust we have developed a new time-dependent neural predictive model for personalized recommendation that balances the long-term evolution of a user's interests with short-term "local coherence". This approach uses parallel recurrent neural networks to capture the evolution of users and items, resulting in a dual factor recommendation model that demonstrates improved performance versus baselines. We have complemented this effort with a new Joint Collaborative Autoencoder framework that learns both user-user and item-item correlations simultaneously, leading to a more robust model and improved top-K recommendation performance. Together this effort has resulted in papers appearing at the 2019 Web Conference (WWW) and 2019 ACM Web Search and Data Mining (WSDM).
- In our second research thrust, our goal is to augment personalized recommendation under fairness-aware constraints. Since recommenders may inherit bias from the training data used to optimize them and from mis-alignment between platform goals and personal preferences, we have been exploring new fairness-aware algorithms that can empower users by enhancing diversity of topics and viewpoints. Specifically, we have proposed a new method to augment tensor-based recommenders with statistical parity-based fairness constraints. The key insight is to extract and then isolate the impact of sensitive features through a tensor factorization approach. Experiments on real-world and synthetic
datasets show that the framework enhances recommendation fairness while preserving recommendation quality in comparison with state-of-the-art alternatives. This effort has resulted in papers appearing at the 2018 ACM Conference on Information and Knowledge Management (CIKM) and the 2nd FATREC Workshop on Responsible Recommendation at RecSys 2018.
- Ziwei Zhu, Jianling Wang, and James Caverlee, Improving Top-K Recommendation via Joint Collaborative Autoencoders (short paper), The International World Wide Web Conference (WWW), 2019. [code]
- Jianling Wang and James Caverlee, Recurrent Recommendation with Local Coherence, 12th ACM International Conference on Web Search and Data Mining (WSDM), 2019.
- Ziwei Zhu, Xia Hu, and James Caverlee, Fairness-Aware Tensor-Based Recommendation, The 27th ACM International Conference on Information and Knowledge Management (CIKM), 2018. [slides] [code]
- Ziwei Zhu, Jianling Wang, Yin Zhang and James Caverlee, Fairness-Aware Recommendation of Information Curators, The 2nd FATREC Workshop on Responsible Recommendation at RecSys, 2018.
This material is based upon work supported by the National Science Foundation under Grant No. 1841138. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.
Date of Last Update: July 2019